Reflection AIOctober 5, 2026

Beam

Reflection AI’s Apache 2.0 open-weight text MoE reasoning model with 501B total and 23B active parameters; Reflection reports 80.9% on SWE-bench Verified and 80.1% on Terminal-Bench 2.1.

Text inputText outputOpen weights

At a glance

Context window
1M tokens
Max output
128K tokens
Text input / output · per 1M tokens
$0.400 / $1.20
Knowledge cutoff
2026-06

Benchmark snapshot

Best result in 6 of 8 categories

View benchmark details (33)

Benchmark details

33 results · Best published score per benchmark, grouped by category

Benchmark · settingScoreRank
Academic
AIME 2026

Reported best (effort unspecified) · Mathematics

97.87 / 46
CritPt

Reported best (effort unspecified) · Scientific Reasoning

16.339 / 134
GPQA Diamond

Reported best (effort unspecified) · Scientific Reasoning

90.546 / 200
HLE

Reported best (effort unspecified) · text-only · Knowledge Exams

36.281 / 151
MysteryMechanism

Unspecified Reasoning (With Tools) · Scientific Reasoning

Source: Vals AI. © 2026 Vals AI. All rights reserved. No affiliation or endorsement.
22.124 / 32
ProofBench v1.1 (Lean 4)

Unspecified Reasoning (With Tools) · Mathematics

Source: Vals AI. © 2026 Vals AI. All rights reserved. No affiliation or endorsement.
18.044 / 54
Coding
AA Cyber Index

Thinking (With Tools) · Repository Engineering

33.714 / 22
DeepSWE

Unspecified Reasoning (With Tools) · Repository Engineering

Shown: Provider-reported. Also reported: Another provider’s table 43.3

44.462 / 73
FrontierCode 1.1

Unspecified Reasoning (With Tools) · Main (100 tasks) · limited comparability · Code Generation & Editing

Cognition, Introducing SWE-2 Flash (November 4, 2026)
22.842 / 51
FrontierCode 2.0

Max (With Tools + Internet) · Repository Engineering

2.934 / 36
IOI (Vals v2)

Unspecified Reasoning (With Tools) · Algorithmic Coding

22.841 / 45
SWE-Atlas-QnA

Unspecified Reasoning (With Tools) · limited comparability · Repository Engineering

34.617 / 17
SWE-bench Multilingual

Unspecified Reasoning (With Tools) · Repository Engineering

78.014 / 48
SWE-Bench Pro V2 Hard

Unspecified Reasoning (With Tools) · Repository Engineering

Shown: Provider-reported. Also reported: Another provider’s table 76.5

77.217 / 21
SWE-bench Verified

Unspecified Reasoning (With Tools) · Repository Engineering

80.99 / 108
Terminal-Bench 2.1

Unspecified Reasoning (With Tools) · Agentic Development

80.138 / 68
Terminal-Bench 4.0

Unspecified Reasoning (With Tools) · vals-mini-swe · limited comparability · Agentic Development

Source: Vals AI, Terminal-Bench 4.0
4.068 / 90
Terminal-Bench 4.1

Normal (With Tools) · Agentic Development

6.444 / 48
Vibe Code Bench v1.1

Unspecified Reasoning (With Tools) · Code Generation & Editing

Source: Vals AI — Vibe Code Bench v1.1. © 2026 Vals AI. All rights reserved. No affiliation or endorsement.
66.453 / 121
Computer Use
MCP-Atlas

Unspecified Reasoning (With Tools) · Tool Orchestration

78.722 / 55
General Evaluation
AA Intelligence Index v5.0.1

Thinking (With Tools) · Capability Indices

17.746 / 62
LiveBench 2.0

Max (task-specific tools) · Cross-capability Suites

Source: LiveBench (livebench.ai)
55.543 / 51
Vals Index v2.2

Reported best (effort unspecified) · Capability Indices

Source: Vals AI — Vals Index v2.2. © 2026 Vals AI. All rights reserved. No affiliation or endorsement.
43.433 / 36
Long Context
AA-LCR

Reported best (effort unspecified) · Long Reasoning

79.315 / 22
LongBench v2

Reported best (effort unspecified) · Long Reasoning

65.54 / 22
Professional
AutomationBench

Unspecified Reasoning (With Tools) · Public set (600 tasks) · limited comparability · Office & Business

37.028 / 64
Legal Research Bench

Unspecified Reasoning (With Tools) · Legal

Source: Vals AI — Legal Research Bench. © 2026 Vals AI. All rights reserved. No affiliation or endorsement.
28.958 / 91
Tax Agent Bench

Unspecified Reasoning (With Tools) · Finance

8.867 / 74
τ³-Banking

Unspecified Reasoning (With Tools) · Service Workflows

38.029 / 120
Reasoning
ARC-AGI-1

Extra-High (No Tools) · Abstract Generalization

74.050 / 87
ARC-AGI-2

Extra-High (No Tools) · Abstract Generalization

23.552 / 80
Research & Search
BrowseComp

Unspecified Reasoning (With Tools) · Fact Finding

77.428 / 61
DeepSearchQA

Unspecified Reasoning (With Tools) · Fact Finding

80.15 / 9
23 more results in Coding, Computer Use, General Evaluation and 4 more

Rank is this model's position among every model with a published result on that leaderboard; the bar shows how far up the field that is. Orange marks first place on leaderboards with at least 5 models. Unranked leaderboards publish a composite or direction-less score.

Back to benchmark snapshot ↑

API pricing

Standard text rates, USD per million tokens · effective 2026-11-13

Input
$0.40
Output
$1.20
Cached input
$0.04

Blended price4:1 input:output · 90% of input tokens cached · same basis as the price ranking

$0.3008/ 1M tokens

Estimate running cost for Beam

Specifications

Provider
Reflection AI
Model type
Reasoning model
Released
October 5, 2026
Total parameters
501B
Active parameters
23B
Architecture
Mixture of experts
Context window
1M tokens
Max output
128K tokens
Knowledge cutoff
2026-06
Input
Text
Output
Text
Licence
Apache 2.0
Commercial use
Commercial use allowed

Capabilities

ReasoningCodingTool useFunction callingStructured outputLong context

Editor’s take

LLMLearner Editorial · Figures as of · AI-assisted

Beam: check the current API limits and the evaluation behind each score

Reflection Beam is an Apache 2.0 open-weight text model with a documented 1M context window. Evaluate its coding results alongside the exact weights, serving configuration and API terms you intend to use.

Beam is Reflection AI's Beam 1.0. Its sparse model activates about 23B parameters from roughly 501B total. That describes computation per token; it does not make full-weight memory or long-context cache requirements disappear.

Use the current API specification. Reflection now lists 1M context and 128K maximum output. The earlier 256K preview limit is no longer the published specification. Your account's model response determines available models and default reasoning/output settings. Record those defaults before measuring latency and cost.

Keep launch results separate from weights evaluations. The catalog retains source-linked published results. Reflection's current model card reports separate BF16, FP8 and NVFP4 runs, so a launch score need not equal a later weights-card score. Compare the same benchmark version, tools and numerics where possible. A repository-fixing result does not establish performance on every coding task.

Choose it for a concrete deployment requirement. Apache 2.0 weights and an API platform give teams different evaluation paths. The catalog lacks a comparable official API quote; that does not establish free use or require self-hosting. Obtain account-specific rates or measure hosting cost at the intended concurrency. Test the same failure cases, count accepted tasks, and include retries and review time.

Use Beam vs DeepSeek V4.1 Flash 1017 for the current paired analysis. It distinguishes model identity, published task evidence and the missing Beam quote. Beam is text-only; evaluate image workflows on an endpoint that explicitly supports them.

Model analysis

Beam (Hugging Face: Beam-1.0-501B-A23B) is Reflection AI’s text-only sparse MoE reasoning model, introduced on October 5, 2026, with 501B total and 23B active parameters for coding and agentic work. In mid-October Reflection published BF16, FP8 and NVFP4 weights on Hugging Face under Apache 2.0; the model card gives a knowledge cutoff of June 30, 2026.

In Reflection’s own table it matches GLM-5.2, not the current Chinese leaders

On Terminal-Bench 2.1 Beam scores 80.1% versus 81.0% for GLM-5.2 and 88.2%, 88.3% and 90.6% for GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. DeepSWE v1.1 is 44.4% (GLM-5.2 44.0%, GLM-5.3 61.0%, DeepSeek V4.1 Flash 74.2%), and HLE without tools is 36.2% versus 42.3% for GLM-5.3 and 46.9% for Kimi K3.

Among US open models it leads clearly

SWE-bench Verified is 80.9% versus 77.6% for Inkling and 70.7% for Nemotron 3 Ultra; SWE-bench Pro v1 is 65.5% versus 54.3% and 46.4%; MCP-Atlas is 78.7% versus 76.0% and 63.1%. All figures are Reflection-run; we found no independent evaluation as of October 19.

Current model and API context: 1M

Checked October 31, 2026: Reflection’s model card and API documentation list 1,048,576 tokens of context and 131,072 tokens of maximum output. The earlier 256K preview limit is no longer the published specification. Reasoning is always enabled, with low, medium, high, xhigh and max efforts; account-specific defaults are returned by the models API. Reflection offers an API platform, while this catalog has no comparable official price quote. Missing pricing does not mean free use or mandatory self-hosting.

Sources: launch blog, Hugging Face model card, model specifications.

Last reviewed 2026-11-01

Editor updates

What changed for this model, noted by LLMLearner Editorial with the source it came from.

  1. Weights released under Apache 2.0

    Reflection published Beam-1.0-501B-A23B on Hugging Face in BF16, FP8 and NVFP4 under Apache 2.0, as promised at launch. The model card states 1M-token context; the preview API remains limited to 256K. Official API pricing is still unpublished.

    Beam-1.0-501B-A23B on Hugging Face

LLMLearner on Beam

Our editorial read, one angle at a time, with the evidence behind each one.

Competitive position

Best US Apache 2.0 coder; a generation behind the Chinese leaders

Our take: Beam matters most to teams that need large open weights under Apache 2.0 from a US lab.

In Reflection's own table it leads other US open models: 80.9% on SWE-bench Verified versus 77.6% for Inkling and 70.7% for Nemotron 3 Ultra, and 78.7% on MCP-Atlas versus 76.0% and 63.1%.

Against current Chinese open models it trails: 80.1% on Terminal-Bench 2.1 versus 88.2% for GLM-5.3 and 90.6% for DeepSeek V4.1 Flash, and 44.4% on DeepSWE v1.1 versus 74.2% for DeepSeek V4.1 Flash. Its closest peer in that table is June's GLM-5.2. BF16 weights are about 1 TB, so self-hosting is a cluster job.

MixedAI-assistedUpdated Oct 18, 2026
2 sources

Beam vs competitors

Comparable results only · record is Beam's wins–losses

Beam benchmark scores against competitors
BenchmarkBeamThis modelvs GLM-5.3Record 0–18vs InklingRecord 18–1
CritPt16.319.15.4
Vibe Code Bench v1.166.478.119.2
τ³-Banking38.050.329.1
Legal Research Bench28.949.028.4
Tax Agent Bench8.873.115.3
AA Intelligence Index v5.0.117.731.712.5
MCP-Atlas78.784.276.0
ProofBench v1.1 (Lean 4)18.049.00.0
FrontierCode 1.122.840.114.0
LiveBench 2.055.560.751.9

Community discussion

Share the API, setup, and tradeoffs you actually tested. Rate the model once; comment as often as you like.

Used this model? Sign in with Google to rate it and share what worked, what did not, and which API or setup you tested.

No community comments yet.